In recent years, the impact of extreme weather has become more and more serious. The losses caused by mangosteen and flying swallows not long ago make us still remember the power of nature. How to better identify extreme weather and do a good job in disaster prevention and mitigation is a huge challenge facing meteorologists.
Recently, in order to help identify potential patterns of extreme weather and predict weather patterns that may harm people’s lives and property, researchers from the University of Berkeley’s Lawrence Laboratory, Oak Ridge Laboratory, and Nvidia have used AI and supercomputers to help The answer is out. They have developed a deep learning system that can identify extreme weather patterns from high-precision climate simulations. This algorithm is expected to help the public predict the weather as soon as possible in the future and respond calmly in the face of natural disasters.
Researchers used Tiramisu and DeepLabv3 neural networks to construct pixel-level masks for extreme weather. In a recently released paper, the researchers described in detail the improvements in the software architecture, including the input process and training algorithm, which effectively improved the large-scale deep learning computing power in supercomputing. The Tiramisu network was deployed on a large scale on 5,300 P100 Nvidia GPUs, achieving 21PF/s computing power and 79% parallel efficiency. At the same time, DeepLabv3 was deployed on 27360 V100GPUs, achieving a speed of 325.8PF/s, while achieving a parallel efficiency of 90.7% under single precision. At the same time, through the efficient use of the FP16 tensor core, the half-precision version of the DeepLabv3+ network can achieve 1.13EF/s peak computing and 999.0 PF/s continuous computing power. These calculations are carried out on the Supercomputing Summit, which is currently ranked number one.
The researchers said that this research has achieved many records. It is not only the first time that deep learning has been used to solve such a large-scale climate segmentation problem in the field of meteorological science, but also the first time that computing power has been extended to Exa ( Ai 10^18) weight level. The computing power of 1.13EF/s is the first time that deep learning has crossed the gate of Exa. It is worth mentioning that this breakthrough also won this year's Gordon Bell Award, which is used to encourage outstanding achievements of the research team in the field of supercomputing.
Let's take a look at the network architecture used by the researchers for pixel-level segmentation of weather conditions.
This work uses an improved DeepLabv3+ network for segmentation of climate data. The encoder part uses the core part of ResNet-50 to extract the high-dimensional features of the input data, and uses the hollow space pyramid pooling to handle larger input resolutions. The decoder is replaced by a unified operation of full resolution in order to generate accurate boundary information. In the above figure, the dark blue box represents the standard convolution operation, the light blue represents the deconvolution operation, and the hole convolution operation is represented by green.
The researchers used CommunityAtmosphere Model (CAM5) to generate 100 years of simulation data, and heuristic algorithms to generate mask labels. The entire data set contains 63k high-resolution samples, which include water vapor, wind, precipitation, temperature and pressure, etc. The segmented regions include tropical cyclones, atmospheric rivers and background three regions.
In order to deploy applications on a large scale in supercomputing, the researchers tested the computational speed of different numbers of GPUs:
The result shows that the current algorithm is very close to the result indicated by the dotted line.
The following figure is the final global prediction mask result. We can see that the predicted result is very close to the black marked result. The blue is the atmospheric long river and the red is the tropical cyclone.
This work was inspired by the previous weather pattern segmentation and successfully applied the Tiramisu and DeepLabV3+ networks to high-resolution multivariable climate data. With the improvement of the loss function, optimization mechanism and network structure, it is quantitative and qualitative. Both achieved excellent results. At the same time, in order to deploy algorithms on a large scale, researchers have established a system-level optimization method, including advanced data storage strategies, optimized throughput and hierarchical communication, so that 27360 VoltaGPUs are used to achieve efficient parallel computing and ultra-high-efficiency computing under the peak node of 4560. With high computing power (peak 1.13EF/s, average 999.0PF/s), this research has expanded the use of open source tools Tensorflow and Horovod, and the developed system has been deployed in Summit supercomputing. This will bring many contributions to the field of deep learning, and will also promote the development of supercomputing platforms.
1KW-8KW Power Inverter
Use double MCU design,the product provides different charge voltages and charge currents to realize charge management for batteries of different types.Its mains supply preferred mode,energy--saving mode and battery preferred mode are all settable,thus making it easy to meet users' different application needs.It has an LCD to show operation status.It is widely applied to families,schools,streets,frontier defense,pasturing areas,industrial equipment,military vehicle--borne equipment,ambulances,police cars,ships,ect.
Nkm Hybrid Inverter With Mppt Charge,Inverter Power Inverter,Hybrid Inverter Charger,Hybrid Grid Tie Inverter
suzhou whaylan new energy technology co., ltd , https://www.whaylan.com